Commits

Breadcrumbs

History for

onmaster

User selector

All users

Datepicker

All time

Commit History

Commits on Dec 11, 2019

• Merge branch 'chong-dev'

  Show description for e87d2c3
  dbolya

  committed
  e87d2c3

  Copy full SHA for e87d2c3

Commits on Dec 6, 2019

• (Hopefully) fixed the inf loss and loss explosion issue by ignoring augmented boxes < 4 px in width and height. Fixes dbolya#222

  dbolya

  committed
  821e830

  Copy full SHA for 821e830

• Release YOLACT++

  chongzhou96

  committed
  ef56a8d

  Copy full SHA for ef56a8d

Commits on Oct 23, 2019

• Optimized training on multiple GPUs.

  dbolya

  committed
  b31107c

  Copy full SHA for b31107c

Commits on Oct 16, 2019

• Merge with master

  dbolya

  committed
  97d2344

  Copy full SHA for 97d2344

• Added a logger so standardize logging and displaying results (so we don't have to dump the commandline every time).

  dbolya

  committed
  12dea3a

  Copy full SHA for 12dea3a

• Added a utility to get info from nvidia-smi

  dbolya

  committed
  25a8998

  Copy full SHA for 25a8998

Commits on Sep 19, 2019

• Added option for Rot90 augmentation (dbolya#153)

  Show description for 02074d5
  

  Agade09

  authored and

  dbolya

  committed
  02074d5

  Copy full SHA for 02074d5

Commits on Sep 17, 2019

• Missing space

  Show description for 706f5db
  Agade09

  authored
  706f5db

  Copy full SHA for 706f5db

• Added option for vertical flipping data augmentation

  Agade09

  committed
  80af768

  Copy full SHA for 80af768

Commits on Jun 12, 2019

• Completely rewrote preserve_aspect_ratio to not need padding and to use varying image sizes during training.

  dbolya

  committed
  bbfae1f

  Copy full SHA for bbfae1f

Commits on May 31, 2019

• Disabled the timer during training because there might be a race condition with multiple GPUs and I wasn't using it anyway.

  dbolya

  committed
  fdd7f23

  Copy full SHA for fdd7f23

Commits on May 19, 2019

• Decoupled frame processing and video playback in --evalvideo and added framerate stabilization for framerates under 30. Now you can use --video_multiframe=some_absurdly_number without worry. We rea…

  Show description for 24de8bc
  dbolya

  committed
  24de8bc

  Copy full SHA for 24de8bc

Commits on May 11, 2019

• Completely overhauled the evaluation pipeline to be much faster, added Pytorch JIT functionality to tons of functions and modules, added multi-GPU support to evalvideo, and removed --no_jit in favo…

  Show description for 8fed726
  dbolya

  committed
  8fed726

  Copy full SHA for 8fed726

Commits on Apr 30, 2019

• Made FastBaseTransform a module and wrapped both it and net with a DataParallel in evalvideo to hopefully support multi-gpu real-time video evaluation.

  dbolya

  committed
  d50616f

  Copy full SHA for d50616f

Commits on Mar 21, 2019

• Traditional NMS implemented!

  dbolya

  committed
  5cbdcc3

  Copy full SHA for 5cbdcc3

Commits on Mar 19, 2019

• Implemented Chong's vectorized cropping and refactored a bunch of the postprocessing code to accommodate it better. As a side effect, not only it the model 3+ms faster, but it's also more accurate!…

  Show description for 89fabd0
  dbolya

  committed
  89fabd0

  Copy full SHA for 89fabd0

Commits on Mar 3, 2019

• Added the ability to and a config to disable certain data augmentation steps.

  dbolya

  committed
  b76da35

  Copy full SHA for b76da35

Commits on Feb 24, 2019

• Sped up evalvideo considerably to the point where video_multiframe actually does something. Gotta go fast.

  dbolya

  committed
  67fc5d8

  Copy full SHA for 67fc5d8

Commits on Feb 15, 2019

• Reworked the implementation of focal loss to be much better and changed it to use softmax. Also reworked the way we use alpha and pi in the same fashion since we're using softmax now. Then I refact…

  Show description for bd2aaad
  dbolya

  committed
  bd2aaad

  Copy full SHA for bd2aaad

• Added reset functionality to MovingAverage and reset them when the delayed config triggers.

  dbolya

  committed
  5824a85

  Copy full SHA for 5824a85

Commits on Nov 10, 2018

• Added Darknet53 finally.

  dbolya

  committed
  a3421aa

  Copy full SHA for a3421aa

Commits on Oct 4, 2018

• Fixed a bug in get_latest that was causing 0 to not be considered latest.

  dbolya

  committed
  5929e52

  Copy full SHA for 5929e52

Commits on Oct 2, 2018

• Finally added the ability to set a crowd iou threshold to count negatives that have a higher crowd iou than that to be considered neutral. I added 2 configs to test it and whether or not I borked a…

  Show description for 367c3e2
  dbolya

  committed
  367c3e2

  Copy full SHA for 367c3e2

Commits on Sep 26, 2018

• Fixed potential issue in the "latest" thing. The weights folder might have files that aren't saved weights.

  dbolya

  committed
  b6a089c

  Copy full SHA for b6a089c

Commits on Sep 25, 2018

• Added the ability to specify "latest" for the train.py --resume option that will find the latest model for the given config.

  dbolya

  committed
  c3f7739

  Copy full SHA for c3f7739

Commits on Aug 31, 2018

• The bug stays.

  dbolya

  committed
  78c5fee

  Copy full SHA for 78c5fee

• Turned off the random channel shuffle (because why).

  dbolya

  committed
  cee7091

  Copy full SHA for cee7091

• Whoops BGR not BRG

  dbolya

  committed
  7b5c9f8

  Copy full SHA for 7b5c9f8

• Added a transform config that allows us to properly preprocess the image for VGG and Resnet differently, fixing a bug where resnet was getting vgg-processed

  dbolya

  committed
  beb805d

  Copy full SHA for beb805d

Commits on Aug 7, 2018

• Fixed up the bbox scaling on the eval script and cleaned some code.

  dbolya

  committed
  a045267

  Copy full SHA for a045267

• Fixed eval augmentations (why did I change that in the first place?)

  dbolya

  committed
  a493439

  Copy full SHA for a493439

Commits on Aug 2, 2018

• Added the ability to change configs as a command-line argument, and added a config for ssd600_resnet101 which is just as it sounds.

  dbolya

  committed
  6637e31

  Copy full SHA for 6637e31

• If preserve_aspect_ratio is set, during eval with --display, the images are now cropped properly and detections outside of the crop are ignored.

  dbolya

  committed
  37105a0

  Copy full SHA for 37105a0

Commits on Aug 1, 2018

• Stomped out the remaining bugs (finally):

  Show description for a02d930
  dbolya

  committed
  a02d930

  Copy full SHA for a02d930

Pagination

Previous

Next